Safety-Critical Control for Dynamic Robots: A Model-based & Data-driven Approach - Koushil Sreenath

In this talk

1. 3:44

   Stability through optimization

   Sreenath explains how control Lyapunov functions describe progress toward a stable state, and how a quadratic program can find a low-effort control input that promotes that progress. The program can also include actuator limits and other constraints.

2. 8:57

   Enforcing safe robot behavior

   Control barrier functions help keep a robot inside a defined safe region by constraining how its safety measure changes over time. Sreenath shows examples including legged robots placing their feet on uncertain terrain and an aerial robot that limits a human pilot’s ability to cross a virtual boundary. He notes that safety depends on choosing a valid barrier function and having a feasible optimization problem.

3. 16:19

   Learning to control Cassie

   Sreenath describes a learned policy for Cassie that uses gait references and robot feedback, trained through imitation learning and simulation randomization. Training used about 400 million samples over roughly a week, followed by checks in a higher-fidelity simulator before hardware deployment. The policy handled walking commands and disturbances, but did not provide stability or safety guarantees.

4. 26:53

   Combining models with learned uncertainty

   One approach trains policies to compensate for errors between a nominal model and the real robot, while retaining CLF and CBF optimization; Sreenath says it still cannot guarantee outcomes because learning may not converge. A Gaussian process approach estimates model error and uses its mean and variance in a second-order cone program to provide probabilistic guarantees. Examples include an inverted pendulum and adaptive cruise control, though the method may not scale well to high-dimensional systems.

5. 46:40

   Limits and practical trade-offs

   In the discussion, Sreenath says safety verification matters as robots interact with people, while noting that more complex systems are harder to verify. He argues that combining known model structure with data can reduce design effort when the model is incomplete. He also points out that perception errors can invalidate safety assumptions, and that time delays are handled in their work through computational optimization and trial and error.